A Visual and VAE Based Hierarchical Indoor Localization Method.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Precise indoor localization is crucial for robotics, augmented reality, and navigation.
• Current visual localization methods often rely on 3D models and can be computationally intensive for large scenes.
• Many deep learning approaches require extensive labeled data, increasing costs and complexity.

Purpose of the Study:

• To develop an unsupervised hierarchical framework for indoor localization and pose estimation.
• To overcome the limitations of labeled data requirements in current deep learning-based image retrieval methods.
• To enable accurate localization using only RGB images.

Main Methods:

• Integration of an unsupervised network variational autoencoder (VAE) with a Structure-from-Motion (SfM) approach.
• Extraction of both global and local features for localization.
• Hierarchical approach using global features for scene-level image retrieval and local features for pose estimation via 2D-3D matches.

Main Results:

• Achieved localization accuracy within 0.16 m and 4° on the 7-Scenes dataset.
• Demonstrated 32.8% accuracy within 5 m and 20° on the Baidu dataset.
• Outperformed advanced methods in terms of precision for indoor visual localization.

Conclusions:

• The proposed unsupervised framework effectively performs indoor localization and pose estimation using only RGB images.
• The integration of VAE and SfM offers a robust solution that bypasses the need for labeled data.
• The method shows significant potential for applications requiring accurate and cost-effective indoor positioning.